[FFmpeg-devel] [PATCH] Fix missing used attribute for inline assembly variables
tejohnson at google.com
Tue Oct 31 18:29:18 EET 2017
It's needed for the same reason the used attribute was already added to the
"static const" variables. For those, when building with just -O2, they
could be removed by optimization since they had local (file) scope, and we
couldn't see the uses in the inline assembly (without the used attribute).
With ThinLTO we have whole program scope, so even though they are
non-static and have global scope we can do dead code elimination on them if
we don't see that they are used.
On Tue, Oct 31, 2017 at 4:19 PM, Michael Niedermayer <michael at niedermayer.cc
> On Tue, Oct 31, 2017 at 03:52:14PM +0000, Thomas Köppe wrote:
> > +Teresa, who drafted the patch initially.
> > On 31 October 2017 at 15:38, Michael Niedermayer <michael at niedermayer.cc
> > wrote:
> > > On Tue, Oct 31, 2017 at 12:16:18PM +0000, Thomas Köppe wrote:
> > > > Variables used in inline assembly need to be marked with
> > > attribute((used)).
> > >
> > > This should not be the case.
> > > Variables referenced through in/out operands should not need that.
> > > Only ones accessed directly bypassing operands should need this
> > >
> > I've added Teresa to the thread, who initially analyzed the problem. (For
> > background on ThinLTO, see here cppcon talk:
> > https://www.youtube.com/watch?v=p9nH2vZ2mNo.)
> > [...]
> > > > - #define DECLARE_ALIGNED(n,t,v) t __attribute__ ((aligned
> > > v
> > > > + #define DECLARE_ALIGNED(n,t,v) t av_used __attribute__
> > > ((aligned (n))) v
> > >
> > > which variables actually are affected/ need this ?
> > >
> > Without this annotation, and when compiling and linking with ThinLTO, I
> > an undefined reference to "ff_h264_cabac_tables", and also to
> > "ff_pw_96", "ff_pw_53",
> > "ff_pw_42", "ff_w1111" and many more.
> i guess these are all accessed directly, like through MANGLE()
> > > Marking all aligned variables as used makes it harder to spot unused
> > > variables leading to accumulation of cruft
> > >
> > I see. Maybe we can annotate only the affected variables more granularly?
> > _______________________________________________
> > ffmpeg-devel mailing list
> > ffmpeg-devel at ffmpeg.org
> > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
> Avoid a single point of failure, be that a person or equipment.
Teresa Johnson | Software Engineer | tejohnson at google.com | 408-460-2413
More information about the ffmpeg-devel